Rate Adaptation of Serial and Parallel Segment Fetching Methods

Transcription

1 Signal Processing: Image Communication ] (]]]]) ]]] ]]] Contents lists available at SciVerse ScienceDirect Signal Processing: Image Communication journal homepage: Rate adaptation for dynamic adaptive streaming over HTTP in content distribution network Chenghao Liu an Imed Bouazizi b Miska M. Hannuksela b# Moncef Gabbouj a# a Department of Signal Processing Tampere University of Technology Finland b Nokia Research Center Tampere Finland article info Keywords: Rate adaptation Dynamic Adaptive Streaming over HTTP DASH Adaptive HTTP Streaming CDN abstract Recently the 3rd Generation Partnership Project (3GPP) and the Moving Picture Experts Group (MPEG) specified Dynamic Adaptive Streaming over HTTP (DASH) to cope with the shortages in progressive HTTP based downloading and Real-time Transport Protocol (RTP) over the User Datagram Protocol (UDP) shortly RTP/UDP based streaming. This paper investigates rate adaptation for the serial segment fetching method and the parallel segment fetching method in Content Distribution Network (CDN). The serial segment fetching method requests and receives segments sequentially whereas the parallel segment fetching method requests media segments in parallel. First a novel rate adaptation metric is presented in this paper which is the ratio of the expected segment fetch time (ESFT) and the measured segment fetch time to detect network congestion and spare network capacity quickly. ESFT represents the optimum segment fetch time determined by the media segment duration multiplied by the number of parallel HTTP threads to deliver media segments and the remaining duration to fetch the next segment to keep a certain amount of media time in the client buffer. Second two novel rate adaptation algorithms are proposed for the serial and the parallel segment fetching methods respectively based on the proposed rate adaptation metric. The proposed rate adaptation algorithms use a step-wise switch-up and a multi-step switch-down strategy upon detecting the spare networks capacity and congestion with the proposed rate adaptation metric. To provide a good convergence in the representation level for DASH in CDN a sliding window is used to measure the latest multiple rate adaptation metrics to determine switch-up. To decide switch-down a rate adaptation metric is used. Each rate adaptation metric represents a reception of a segment/portion of a segment which can be fetched from the different edge servers in CDN hence it can be used to estimate the corresponding edge server bandwidth. To avoid buffer overflow due to a slight mismatch in the optimum representation level and bandwidth an idling method is used to idle a given duration before sending the next segment. In order to solve the fairness between different clients who compete for bandwidth the prioritized optimum segment fetch time is assigned to the newly joined clients. The proposed rate adaptation method does not require any transport layer information which is not available at the application layer without cross layer communication. Simulation results show that the proposed rate adaptation algorithms for the serial and the parallel segment fetching methods quickly adapt the media bitrate to match the end-to-end network capacity provide an advanced convergence and fairness between different clients and also effectively control buffer underflow and overflow for DASH in CDN. The reported simulation results demonstrate that the parallel rate adaptation outperforms n Corresponding author. Tel.: þ address: chenghao.liu@tut.fi (C. Liu). # EURASIP member /$ - see front matter & 2011 Elsevier B.V. All rights reserved. doi: /j.image

2 2 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] the serial DASH rate adaptation algorithm with respect to achievable media bitrates while the serial rate adaptation is superior to the parallel DASH with respect to the convergence and buffer underflow frequency. & 2011 Elsevier B.V. All rights reserved. 1. Introduction Recently Hypertext Transfer Protocol (HTTP) has been widely used for the delivery of real-time multimedia content over the Internet such as in video streaming applications. Unlike the use of Real-time Transport Protocol (RTP) over User Datagram Protocol (UDP) HTTP [1] is easy to configure and is typically granted traversal of firewalls and network address translators (NATs) which makes it attractive for multimedia streaming applications. Krasic et al. [2] Wang et al. [3] and Kim and Ammar[4] reported that short-term transmission rate variation in HTTP/TCP can be smoothed out by receiver side buffering. Thus commercial solutions such as Microsoft Smooth Streaming [5] and Adobe Dynamic Streaming [6] have been launched as well as standardization projects have been carried out. Adaptive HTTP streaming (AHS) was first standardized in Release 9 of 3GPP packetswitched streaming (PSS) service [7]. MPEG took 3GPP AHS Release 9 as a starting point for its newly published MPEG DASH standard [8]. 3GPP continued to work on adaptive HTTP streaming in communication with MPEG and recently published the 3GP-DASH (Dynamic Adaptive Streaming over HTTP) [9]. MPEG DASH and 3GP-DASH are technically similar and are therefore collectively referred to as DASH in this paper. In DASH the client continuously requests and receives small segments of multimedia content denoted as media segments. To adapt the media bitrate to the varying network bandwidth DASH allows clients to request media segments from different representations each of which represents a specific media bitrate. A DASH client might suffer from frequent interruptions and sub-optimum media bitrates without an efficient rate adaptation algorithm. A media bitrate higher than the sharable bandwidth would cause network congestion. As DASH typically uses Transmission Control Protocol (TCP) as transport layer protocol network congestion causes TCP congestion control mechanism to become active which may further result in a dramatic decrease in throughput. Hence buffered media data can be drained-up much faster in DASH compared to the traditional RTP/UDP-based streaming in case of network congestion. On the other hand i.e. when the media bitrate is lower than the sharable bandwidth the media quality does not reach the optimum allowed by the available bandwidth. Many of the Internet s real-time video services use a progressive download approach and hence suffer from the above mentioned problems such as frequent playback interruption and sub-optimum streaming quality. Many video services offer a set of pre-defined quality levels of a video clip to users for manual a-priori selection. If the bitrate of the selected representation turns out to be higher than the available end-to-end bandwidth then the user will most probably experience playback interruptions and re-buffering events due to buffer underflow. Otherwise if the bitrate of the representation is lower than the available network bandwidth then the user will consume the content at a sub-optimal quality. Since the client will download content faster than the playout rate it could result in inefficient use of available bandwidth if the user stops watching the content. An efficient rate adaptation algorithm is required for DASH to solve the above problems specifically the frequent interruption in playback and sub-optimum streaming quality. However these problems become even more challenging because of the difficulties in differentiating between the short-term throughput variation incurred by the congestion control and the average throughput changes due to more persistent bandwidth changes. Furthermore the rate adaptation algorithm of DASH should take into account the infrastructure of the networks. The research issues of the rate adaptation for DASH include the following aspects. First the rate adaptation method must deploy a metric to determine if the bitrate of a specific representation matches the available end-toend bandwidth. This metric is expected to distinguish between persistent throughput variation due to network bandwidth changes and the short-term throughput dynamics attributable to the TCP congestion control. In addition the metric should identify network congestion and spare network bandwidth fast enough in order to react promptly and reach the optimum representation level as soon as possible wherein the higher representation level represents the higher level of media bitrates and the lower representation level represents the lower level of media bitrates. Second the rate adaptation algorithm should prevent frequent hopping between media representations because frequent changes in the perceived media quality are likely to be annoying. In segment delivery over distributed networks such as content distribution network (CDN) preventing too frequent representation switching becomes more challenging as segments can be transmitted through different routes having different bandwidths. Thirdly the rate adaptation algorithm should avoid client buffer underflow and overflow. Buffer underflow causes playback interruptions and overflow can result in bandwidth waste. Fourthly the rate adaptation algorithm for DASH should provide good fairness between different DASH clients which compete for available bandwidth in the network. Finally the media segment duration should be set appropriately in order to smooth the short-term HTTP/TCP throughput variation and to provide low enough adaptation latencies. Two strategies for issuing HTTP requests in DASH have been proposed namely the serial segment fetching method [10] and the parallel segment fetching method [11]. In the former segments are requested and received one after another whereas in the latter media segments are requested in parallel using several TCP connections. This paper presents novel rate adaptation algorithms for both the serial and the parallel segment fetching methods.

3 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 3 The main contributions of this paper include the following two aspects. First we propose a novel rate adaptation metric for DASH. The proposed rate adaptation metric compares the expected segment fetch time ðesftþ with the measured segment fetch time ðsftþ to determine if the media bitrate matches with the current bandwidth. The ESFT is the optimum segment fetch time which is estimated each time before requesting a segment based on (a) the segment duration multiplied by the number of parallel segment fetching threads (b) the remaining time period to fetch the next segment. The former is determined by the scheduling method of the serial and the parallel segment fetching methods whereas the latter is decided by the real-time buffered media time. The rate adaptation metrics not only take into account the ideal segment fetch time of the serial and the parallel segment fetching methods but also consider the buffered media time. Hence the proposed rate adaptation metrics can provide convergence in the representation level and prevent buffer underflow. Second we propose two novel rate adaptation algorithms for the serial and the parallel segment fetching methods in CDN respectively. The proposed rate adaptation algorithms for the serial segment fetching method uses a single rate adaptation metric to determine a switch up/down of representation levels. The proposed rate adaptation algorithm for parallel segment fetching method deploys a sliding window to select the minimum rate adaptation metrics among multiple metrics. The minimum rate adaptation metrics is used to determine the switch-up/ switch-down in order to match the media bitrates to the minimum bandwidth in the distributed edge servers of CDN. Thus the sliding window strategy of the proposed rate adaptation algorithm for the parallel segment fetching method can efficiently prevent frequent hopping between representation levels. Furthermore the proposed rate adaptation algorithms give a priority for newly joined DASH clients in order to share the available bandwidth fairly among DASH clients. In case of congestion DASH clients consuming a higher representation level will quickly switch down to a lower level firstly in order to leave room for DASH clients having a lower level using the proposed rate adaptation algorithm. Furthermore in order to efficiently use network bandwidth and memory resources for the users a method of determining the idle time before sending an HTTP GET request for the next segment is presented thus limiting the maximum amount of pre-fetching media. The rest of this paper is organized as follows. Section 2 presents the related work to describe the literature works relating to the rate adaptation for DASH. Section 3 gives a brief overview of the terminology of DASH and DASH infrastructure using a CDN. The serial and the parallel segment fetching methods are presented in Section 4. In Section 5 a novel rate adaptation metric for DASH is presented. Sections 6 and 7 provide further details on the two novel rate adaptation algorithms for the serial and the parallel segment fetching methods in CDN respectively. The simulation results of the proposed rate adaptation algorithms for the serial and the parallel segment fetching methods in CDN are provided in Section 8. Finally Section 9 concludes the paper. 2. Related work In the last decade several research works have been conducted in the rate adaptation for media streaming over TCP called TCP-based streaming in this paper. In the scenario of the TCP-based media streaming most of the research works focused on sender-driven rate adaptation. For example Lam et al. [12] proposed a method to estimate client-side buffer occupancy in the server and to adapt media bitrates to maintain the client-side buffer occupancy aboveacertainthreshold. Rate adaptation for DASH is a relatively new research topic and thus only few research works have been carried out in this field. In DASH decisions to adapt media bitrates are made at the client side where the client buffer occupancy is known and can be used directly for the rate adaptation. For example in one of the earliest papers on adaptive streaming over HTTP by Färber et al. [13] rate adaptation decisions were made on the basis of media duration in the client buffer. However client buffer occupancy based rate adaptation for DASH has the following drawbacks. First this method can cause hopping in the representation level. As the choices of representation levels are limited it is difficult to match the optimum media bitrate to the network capacity. Typically the optimum media bitrates should be the highest bitrate which is lower than the bandwidth. In such a scenario the client buffer occupancy which can be measured as buffered media time will dramatically increase even if the rate adaptation reaches the optimum media bitrate. Hence the rate adaptation cannot converge to the optimum representation level as the client buffer occupancy will change even when the rate adaptation reaches the optimum media bitrates. Second the client buffer based rate adaptation can result in unfairness between different clients which compete for bandwidth. For example some clients can switch to a high-bitrate representation if adequate bandwidth is available while other clients might not be able to fetch equally high-bitrate representation if the networks are already on the stage of saturation when they start streaming. Efficient rate adaptation algorithm should provide a fair allocation of the network bandwidth to clients. Kuschnig et al. [14] presented a method of priority and deadline driven rate adaptation method for scalable video streaming over TCP using bandwidth estimation and client buffer estimation at the server. To stabilize TCP throughput over a given period of time and decrease quality fluctuation relatively large groups of pictures (GOPs) are used to estimate TCP throughput by deploying the application layer based and TCP stack based bandwidth estimation techniques. As pointed out in [14] the accuracy of the bandwidth estimation method suffers from the large GOP which is coped with the priority and deadline driven rate adaptation method proposed in the same paper. Recently Akhshabi et al. [15] reported experimental evaluation of the rate adaptation in DASH. In [15] the experiment is performed using the Microsoft smooth streaming Netflix player and Adobe OSMF player to evaluate the internal rate adaptation algorithms with certain experimental methodology. In our previous work [10] a rate adaptation algorithm for the serial segment fetching method was presented. In

4 4 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] [10] the ratio of the segment duration and the latest single segment fetch time was used to detect network congestion and spare network capacity. However the rate adaptation algorithm in [10] has the following shortcomings. First the rate adaptation method was developed for the case where segments are requested and received sequentially and hence ESFT is determined in [10] to be equal to the media segment duration ðmsdþ. However for the parallel segment fetching method where the number of simultaneous HTTP transactions can be larger than one ESFT should also be proportional to the number of parallel HTTP threads. Second the buffered media time status of the client buffer is not taken into account in the rate adaptation metric proposed in [10] which causes buffer underflow due to the lower buffered media time. To prevent buffer underflow the buffered media time should be used as the metric for determining the next segment media bitrates. Third the rate adaptation method compares the ratio of the segment duration by the latest received single segment fetch time with the switch-up and switch down thresholds to adapt the representation level for the serial segment fetching method. In the parallel segment fetching method multiple segments are received simultaneously hence multiple segment fetch times should be used to decide whether or not to perform switching-up in order to improve convergence to a single representation level when the networks resource capacity is not changed severely. Fourth the fairness issue between different DASH clients has not been discussed in [10]. DASH may employ the current web distribution network infrastructure e.g. proxy caches and CDN. Currently CDNs are widely used to deliver media and web data to end users through globally distributed edge servers. CDNs such as Akamai deploy Domain Name System (DNS) based request redirection technique to redirect a client request to an appropriate edge server among multiple edge servers with the predefined policies. Each client request is used to fetch a media segment in DASH. The detailed request redirection mechanism and the redirection frequency are described in Section 3. Scheduling of HTTP GET requests for segment fetching plays a critical role in achievable streaming bitrates for DASH in CDN. However the specific scheduling method of requesting segments has not been specified in DASH. Only the informative instructions of the client behavior are provided in both 3GP-DASH and MPEG-DASH. One common method is the serial segment fetching method which requests and receives segments one after another. For the DASH in CDN the serial segment fetching method might not yield the optimum streaming service to DASH clients since it only operates one HTTP thread at a time to deliver media segments over the distributed edge servers in CDN which results in inefficient usage of the distributed network resources. The parallel segment fetching method can occupy network resources of CDN more exhaustively as multiple HTTP threads are used to deliver the media segments in parallel over distributed networks. In our previous work [11] a parallel adaptive HTTP streaming method was presented. The method schedules HTTP GET segment requests in parallel at the DASH client for efficiently utilizing the distributed networks resources and improving the achievable streaming quality. Specifically our previous paper presented a method which (a) enables a receiver to request multiple segments in parallel (b) provides a solution to maintain a limited number of HTTP sessions for receiving segments in parallel and to determine when to start a new HTTP session to request the next segment and (c) can adapt the received media bitrate while receiving previously requested segments. For the rate adaptation algorithm [11] deployed a simple rate adaptation algorithm which uses the buffered media time to determine the media bitrates hence results in similar problems as described above for [13]. Since the focus was on improving achievable representation level the simulation results in [11] were limited in reporting the average representation level for a single DASH client user. In this paper we investigate the rate adaptation for the serial and parallel segment fetching in CDN. In a multiple TCP connection based media streaming requesting a set of media chunks using multiple TCP streams was presented in [1617]. Multiple TCP connection based media streaming was previously presented in [18]. The work in [1617] extended multiple TCP connections based media streaming to the requestresponse based internet video streaming. In the same papers the authors presented analytical models for the total achievable TCP throughputs for the request-response based TCP streaming using multiple TCP connections. Although this paper uses the parallel HTTP requests the aim and the underlying principles are totally different from those in [1617] where the aim of multiple TCP connections was to provide resilience against short-term insufficient bandwidth using multiple TCP connections for a streaming application. In contrast the target in our previous paper [11] was to fully utilize the distributed network resources for example CDN to improve HTTP- Streaming quality at the client. The underlying principle of the works [16 18] was that by allocating a certain bottleneck bandwidth to multiple TCP streams the TCP throughput variation can be smoothed out and the achievable TCP throughput can be increased to some degree as multiple TCP streams do not observe packet losses at the same time. In contrast [11] is inspired by the distributed infrastructure of CDN. Evensen et al. [19] presented a method of deploying multiple heterogeneous interfaces to improve streaming quality. They proposed to divide a segment into multiple sub-segments dynamically and request sub-segments using distributed requests over multiple heterogeneous interfaces simultaneously. As [19] uses sub-segments several techniques for partitioning segments into subsegments were studied and the paper proposed a dynamic and static sub-segment partitioning method. In contrast our previous paper [11] used segment as a fetching unit which makes the parallel segments fetching method simpler compared to the parallel sub-segment fetching method [19]. The underlying principle of paper [11] is the efficient utilization of the distributed edge servers in CDN. On the contrary [19] considered a device using multiple heterogeneous network interfaces such as a smart phone equipped with the ability to transfer data through WLAN and HSDPA networks simultaneously.

5 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 5 3. Terminology and architecture of DASH in CDN In this section we shall first explain the terms and definition of DASH as specified in 3GPP PSS. Then the end-to-end systems of DASH including the DASH content provider DASH server and DASH client are described. In addition the media distribution architecture in CDN is presented. In DASH the media presentation description (MPD) provides the necessary information for clients to establish a dynamic adaptive streaming over HTTP. The MPD contains the metadata such as an HTTP-URL of each segment to make GET segment request. The segment contains a certain duration of media data and metadata to decode and present the included media content. Each representation consists of multiple media segments and represents the coding choice such as encoded bitrate spatial and temporal resolution etc. The representation is identified through a unique identifier (ID). The client can obtain the available number of representations and the corresponding properties by accessing the MPD. Fig. 1 shows a media distribution architecture using DASH in CDN. An end-to-end DASH system consists of a server clients and a content provider. DASH content preparation prepares the DASH compliant media content in an offline segment creation mode or in a reactive segment creation. In the offline segment creation mode the MPD and the media segments for the representations are created before providing client access to the content. In the reactive segment creation mode the content provider constructs the media segments according to the received HTTP GET request from a media file such as an MP4 file which contains multiple representations. In the second mode for example the client request may contain the segment start time segment duration and representation level. The content provider firstly parses the HTTP request and encapsulates part of an MP4 file to a segment as specified in the HTTP request. A DASH content provider sends DASH compliant media data to a DASH server for streaming. The DASH server can be a standard web server to provide dynamic adaptive streaming over HTTP to DASH clients. The DASH client firstly accesses MPD to obtain the representation and the address to locate each segment for making HTTP GET segment request. The client continuously requests and receives the segments from the DASH server. DASH further enables the client to request media segments from different representations for reacting to varying DASH content preparation DASH server CDN Edge servers DNS CDN server DASH-clients group Fig. 1. System for Dynamic Adaptive Streaming over HTTP (DASH) in CDN. network resources. The rate adaptation can occur each time before requesting a segment. As DASH facilitates the use of the current web replication and caching infrastructure CDN is used in this paper to deliver media segments to DASH clients. CDN has achieved great success in delivering web and media data. The infrastructure of CDN such as Akamai consists of a set of edge servers which deliver content to clients as shown in Fig. 1. CDN can redirect a client request such as a GET media segment request to an appropriate edge server among multiple edge servers. The edge server selection method especially the frequencies of switching edge server affects the performance of serial and parallel segment fetching methods. For locating the edge server from which an end user receives a segment the DNS request redirection is used to redirect a client request to an appropriate edge server in CDN such as Akamai [20]. The frequencies of switching the edge server depends on the deployed metrics by the DNS based request redirection. To select an edge server from multiple edge server candidates the metrics of proximity content caching possibility and edge server load can be used in a DNS request redirection system [20]. Proximity indicates the distance between the edge server and the client and a close-by edge server will be selected. Edge server load is measured as the bandwidth and processing loads of the edge servers. The edge server load balancing algorithm distributes traffics among edge servers targeting a uniform load distribution at the edge servers. Basically the proximity based edge server selection method results in the infrequent switching between edge servers since the distance between the edge server and the client is relatively static. In contrast the load based edge server selection method causes frequent switching when balancing the load between different edge servers in order to improve the overall performance. To achieve improved load balancing short time-to-live (TTL) is typically used in the DNS based request redirection also called DNS resolution. For DNS based request resolution [21] the authoritative DNS name server and local DNS server cooperate in a hierarchical way. The local DNS server sends a DNS request to an authoritative DNS server which resolves the CDN server name to the IP address of the edge server. In the IP address mapping received from authoritative DNS server a time-to-live (TTL) is typically contained for the mapping. When the local DNS server queries the authoritative DNS server the local DNS server will cache the content record for the time specified in the TTL. If the local DNS server receives a query about the caching name server for the mentioned record within the TTL the local DNS server resolves the edge server IP address with the cached content recoding instead of querying it from the authoritative name server. Krishnamurthy et al. [21] reported that generally a low TTL is assigned in order to assist the DNS based load balancing among the distributed edge servers. The authors claimed that the local DNS servers are forced to process frequent DNS lookups by assigning a small DNS TTL for the content record. Therefore CDN can obtain control for utilizing the distributed edge servers efficiently by taking advantage of

6 6 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] the load balancing algorithm. As an observation of the real CDN request redirection they reported that the Akamai and Adero Digital Island Speedera and Fasttide assigned and 230 s of TTL respectively. Recently Adhikari et al. [22] presented research results of uncovering load balancing and server selection strategies deployed in Youtube using a reverse engineering based methodology. They reported that Youtube deploys the static load sharing using hash based mechanism a semi-dynamic approach using location aware DNS resolutions and dynamic load-sharing using HTTP redirections. In the first strategy Youtube maps video-id to a unique hostname in each of the namespace in the hierarchical DNS based host namespaces. In the second strategy the proximity information is used in request redirection according to the normal hours and busy hours. In the third strategy Youtube uses local intra-tier loadsharing before resorting to inter-tier load-sharing to further balance the load among different edge servers. The research results in [22] showed that the load balancing is deployed in practical applications such as Youtube to smooth the un-even load distribution caused by the combination of spontaneous video demands and dynamic video popularity. The inter-tier load-sharing infers that frequent edge server switching is possible to balance the load in the distributed edge servers. Therefore redirecting a HTTP request for fetching a specific segment to a specific edge server depends on the load balancing algorithm of DNS based request redirection and TTL value used in CDN. In addition the dynamic caching status of media segments and the varying network resources further affect the load balancing algorithm to select an appropriate edge server. Several factors including load balancing algorithm TTL value used and dynamic caching and network resources together determine the switching between different edge servers for the consecutive requests from the same client. Specifying such frequency is out of the scope of this paper. To focused on rate adaptation algorithm of DASH the proposed rate adaptation algorithm is presented as transparent of the edge server switching frequencies. In the content outsourcing pull-based and push-based outsourcing can be utilized [23]. In the pull-based content outsourcing the client request is redirected to the selected edge server as discussed according to the above policies using the DNS redirection or URL rewriting methods. The pull-based outsourcing can be classified into uncooperative and cooperative outsourcing. For both uncooperative and cooperative pull-based outsourcing if the requested segment is cached in the selected edge server then the client fetches the segment directly from the edge server. In the case of a cache miss the uncooperative outsourcing redirects the request either to the CDN server such as Akamai server or to the original server. For the cooperative outsourcing the edge servers cooperate with each other to response to the request from a client in case of a cache miss. The push-based content outsourcing distributes the contents to the edge servers proactively. As the push-based outsourcing is on the theoretical study stage this paper focuses on the pull-based outsourcing. For the uncooperative and cooperative outsourcing we focus on the former strategy as most of CDN providers such as Akamai deploy this method [23]. Thanks to the distributed resources offered by CDNs a DASH service over a CDN can improve the scalability in streaming clients. However the achievable bandwidth for delivery of media segments is still limited. In media delivery over CDN congestion might not only occur in the last mile of media delivery network e.g. the link from the Internet service provider to the end user equipment but also in the distributed edge servers. Furthermore more and more end users are connected to the Internet using a high-bandwidth access link. Hence efficient utilization of the distributed edge servers in CDNs is one of the key factors to provide high quality media in DASH. Different segments can be delivered to DASH clients through different edge servers in CDN with request redirection and load balancing algorithm. Therefore a parallel segment fetching method was proposed in our previous paper [11] to utilize the limited bandwidth in the distributed edge servers more efficiently. As shown in Fig. 1 the scenario of multiple DASH clients which compete for the distributed bandwidths in the edge servers is discussed in this paper. To investigate the uncooperative pull based content outsourcing in the scenario of multiple DASH clients different DASH clients request a different media clip. By requesting different media clips each GET segment request is sent to the CDN server which then pulls the segment from the initial server. 4. Serial and parallel segment fetching methods In this section the serial and the parallel segment fetching methods are briefly described [1011]. A summary of the terms and a definition of the symbols appearing in this paper are shown in Table 1. The serial segment fetching method [10] is a common method to request a segment where segments are requested and received one after another. Fig. 2 shows an example of requesting segments serially. The horizontal axis denotes the time of requesting and receiving segments and the period of time to receive segment #i (rec segment #i) is depicted in Fig. 2. Note that in general the next segment need not be requested immediately after completing the reception of the previous segment but rather segment requests can be scheduled according to the buffer occupancy and the rate adaptation strategy of the client. The detailed flowchart of the serial segment fetching method will not be discussed further because of its simplicity. In the parallel segment fetching method [11] segments are requested in parallel. Fig. 3 shows an example of the parallel segment fetching method. In addition the time axis is depicted to scale the time period to request and receive segments. As shown in Fig. 3 the parallel segment fetching method maintains a certain number of parallel HTTP threads to request segments in parallel e.g. two HTTP threads are used in this example. The purple and dark blue bars denote the receiving segments over the first and the second HTTP threads respectively. The purple and dark blue lines indicate the time axis to request and receive the media segments respectively. The parallel segment

7 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 7 Table 1 Terms and definition of symbols. Notation Definition A Slope of the linear function of m k used in Eq. (2) b rc Encoded media bitrate of the current representation level r i b ri Media bitrate of the representation level wherein r 0 and r N 1 denote the lowest and the highest representation b max Media bitrate of the highest representation level b min Media bitrates of the lowest representation level bmt c Current buffered media time bmt min Minimum buffered media time used to determine t id c The current representation level DASH Dynamic adaptive streaming over HTTP d Ratio of the minimum received portions duration and segment duration to determine s 0 for smoothing the short-term TCP variation e d Switch-down factor e c d Switch-down factor calculated by the current representation level e min d Minimum switch-down factor calculated by all representation levels EPFT Expected portions fetch time e u Switch-up factor e c u Switch-up factor calculated by the current representation level e max u Max switch-up factor calculated by all representation levels ESFT Expected segment fetch time u k The ratio of the duration of the received portion and the duration of the full segment used for scheduling a new parallel HTTP request ISFT Ideal segment fetch time k Parallel index wherein the latest segment has the index 0 and the earliest segment has the index n p 1 l k Threshold of j k to schedule a new parallel HTTP request l 0 First term of the linear function of m k used in Eq. (2) MPD Media portion duration denoting the playback duration contained in the received portions MPD k Received media portion duration of a segment with the parallel index k MSD Media segment duration denoting the playback duration contained in a segment MSD k Media segment duration contained in the segment with the parallel index k n oo Number of the out order received segments n u oo Upper limit of n oo n p Number of parallel HTTP threads to deliver media segments n u p Upper limit of HTTP threads number to deliver media segments in parallel PFT Portion fetch time denoting the spent time for fetching the portions of a segment PFTM Portion fetch time based rate adaptation metric RAM s Rate adaptation metric including PFTM and SFTM with the rate adaptation index t RAM s0 The rate adaptation metric with the rate adaptation index t 0 which is determined as Eq. (21) rec segment spent time to fetch segment #i #i q RSFT s factor r i Representation level i RSFT Remaining segment fetch time RSFT p The priority remaining segment fetch time for parallel segment fetching method which will be used in the beginning phase of DASH RSFT s The priority remaining segment fetch time for serial segment fetching method which will be used in the beginning phase of DASH SFT Segment fetch time SFTM Segment fetch time based rate adaptation metric s Rate adaptation index of the segment/portions which starts from t 0 and ends at t 0 þw 1 s 0 The staring rate adaptation index of the segment/chunk TBMT Target buffered media time t id Idle period of time t_req i Time to request a segment #i ts i Playback timestamp of the first frame of a requesting segment # i ts ns Playback timestamp of the first frame of the next segment ts 0 The current playback time stamp at the time instant of requesting the next segment u p Upper limit of parallel HTTP threads w Rate adaptation window size which determines the account of RAM used in the rate adaptation Fig. 2. An example of serial segment fetching method. fetching method requests segment iþ1 att_req i þ 1 using the second HTTP thread when segment i is still being received using the first HTTP thread by the client. From time t_req i þ 1 to the time to receive the last byte of segment #i the two segments i.e. segment #i 1andsegment#iþ1 are received in parallel.

8 8 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] Fig. 3. An example of parallel segment fetching method. The scheduling method of HTTP GET segment requests in parallel is described as follows. The number of active parallel HTTP threads ðn p Þ is compared with predefined upper limit ðn u pþ to limit the maximum number of parallel HTTP threads to deliver media segments in parallel. We recommend to set n u p as 2 since two parallel HTTP requests already can improve the achievable media bitrates according to the simulation results [11]. In addition a large value of n u p may increase the possibility of receiving segments out of order which may further cause buffer drain-up also called buffer underflow. Each segment received in parallel is identified with a parallel index k wherein the last requested parallel index is equal to 0 and the earliest requested parallel index is equal to n p 1. When requesting a new segment in parallel all parallel indexes are updated in turn as mentioned above. The parallel index is different from the segment IDs marked as #i in Fig. 3. In each received segment the portion of received media data is referred to as received portion and media duration contained in a received portion is denoted as MPD. The ratio of the duration of the received portion and the duration of the full segment j k is derived as follows j k ¼ MPD k =MSD k 0rkrn p 1 ð1þ The DASH client compares j k with the corresponding threshold m k for each parallel received segment k from index 0 to index n p 1 to determine if a new HTTP thread needs to start to fetch the next segment. If j k is larger than or equal to m k for all k from 0 to n p 1 then a new request may be sent and the client checks if the remaining conditions to send the next HTTP request are met. Otherwise (if j k is lower than m k for any k in the interval f0; n p 1g) the new request will not be sent. The remaining conditions to send the next request will be presented in the following paragraphs. To receive the earlier segment in time a linear function m k is used as m k ¼ Akþm 0 0om k o1 ð2þ In case m k is larger than or equal to 1 it is set as 1. In this paper A is set as 0.2 as in our previous paper [11]. m 0 can be set as 1/n p. Larger m 0 and A will assist to successfully receive the earlier requested segments but deduce the efficiency in improving the achievable media bitrates. In addition the requested segments may be received out of order. To assist receiving the earlier requested segments in time the client does not request a new segment in parallel in case the gap between the latest requested segment ID and the earliest requested segment ID is equal to or larger than its upper limit ðn u ooþ. The smaller n u oo will facilitate to receive the earlier requested segments in time but will limit the capability of parallel segment requesting method with respect to improving receiving media bitrates. In this paper n u oo is set as 5. If n p is smaller than n u p n oo is smaller than n u oo and j k is larger than or equal to m k for all k from 0 to n p 1 and a new HTTP request is sent immediately or after a certain idling period derived from the buffered media time. The buffered media time is limited for preventing bandwidth waste due to the fact that a client may stop viewing a certain media clip before playing out the whole downloaded media data. The idling period between two consecutive requests is used to limit the maximum amount of buffered media time. The idling period must not be too long so that the buffered media time can smooth out the most severe bandwidth decrease i.e. from the highest bandwidth to the lowest bandwidth. Therefore the idle time period is calculated as t id ¼ bmt c bmt min MSDUb max =b min ð3þ If t id is greater than 0 then the client idles t id amount of time before requesting the next segment. Otherwise (if t id is less than or equal to 0) the client requests a new segment immediately. In case of a small MSDUb max =b min bmt min should be set to a relatively large value to increase the condition of bmt c for idling and to prevent buffer underflow due to sudden bandwidth decrease. In case of a large MSDUb max =b min bmt min can be set as a relatively small value because bmt c for idling becomes large enough. bmt min is set to 0 in this paper since MSDðb max =b min Þ is already large enough with the selected media bitrates of the representations in the simulation to prevent buffer underflow. For requesting the next segment the rate adaptation module should be performed to determine the representation level of the next segment. Then n p is increased by one when a new HTTP thread is started to request the next segment in parallel. To conclude the discussion above the DASH client does not request a new segment and continues to receive the previously requested segments in the following three cases. First the number of segments being received equals to the limit of the number of parallel HTTP threads i.e. n p is equal to n u p. Second j k is smaller than m k in one of the segments being received. Third the idle period is greater than 0. With the chunked transferring mode of HTTP the DASH server can split a media segment into small portions and transmit a segment as several portions. The client can receive

9 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 9 a segment portion-by-portion until receiving the whole segment. When new portions of a segment are received by the client the client updates the duration of the received portion ðmpd k Þ for the parallel received segment. If the whole segment is received by the client then n p is decreased by one. Finally if all segments are received or the user ends the current DASH then the current parallel segment fetching is terminated. The detailed scheduling algorithm is presented in the following pseudocode. shortly called as portion in this paper is received. As the reception data rate of the latest received portion can represent the prevailing network bandwidth the measured portion fetch time ðpftþ is compared with the corresponding expected portion fetch time ðepftþ which is used as another rate adaptation metric denoted as PFT rate adaptation metric ðpftmþ. Specifically SFT denotes a period of time from the time instant of sending a HTTP GET segment request to the Scheduling algorithm for HTTP GET segment requests for the parallel segment fetching method in CDN [11] Initialize the current representation level i as 0 Send HTTP GET request to fetch the first segment immediately While User has not stopped the current DASH and more segments are available for fetching Wait for an event If Received portions of the requested segments If The received portions of a segment having parallel index k Update MPD k If Received a whole segment n p ¼ n p 1 Else If Timeout of the scheduled event to send the next HTTP GET request Send the next HTTP GET request immediately (after processing the rate adaptation module) n p ¼ n p þ1 If n p on u p and noo onu oo k ¼ 0 While kon p If j k Zm k k ¼ kþ1 Else Break End while If k ¼ n p Calculate the idle period t id as Eq. (3) If t id 40 Schedule an event to send the next HTTP GET request after idling t id amount of time Else Send the next HTTP GET request immediately (after processing the rate adaptation module) n p ¼ n p þ1 End while 5. Proposed rate adaptation metric This section presents the proposed rate adaptation metric for DASH called segment fetch time ðsftþ rate adaptation metric ðsftmþ which compares ESFT with the measured SFT. The former represents the optimum segment fetch time and the latter denotes the measured segment fetch time. It is well known that the instantaneous TCP transmission rate is dynamically changing hence it is not feasible to measure the streaming media bitrate capacity using the instantaneous HTTP/TCP reception rate. Instead a DASH client estimates the optimum segment fetch time i.e. ESFT each time before requesting a segment. After receiving a segment the estimated ESFT is compared with the measured SFT to determine if the media bitrate of the current representation matches the available end-to-end bandwidth capacity. In the parallel segment fetching method the HTTP GET request can be sent when a portion of the previous segment instant of receiving the last byte of the requested segment; while the PFT denotes a period of time from the time instant of sending a HTTP GET segment request to the instant of receiving the last byte of the portion of the requested segment. To smooth out short-term variations in HTTP/TCP reception rate the measuring period should be selected appropriately. Typically a longer period is capable of producing a smoother throughput measurement. However a longer period will also result in a slower rate adaptation behavior. Based on the theoretical analysis of smoothing short-term TCP throughput reported in our previous paper [24] and experimental results reported in the same paper a measuring duration around 7 10 s is sufficient to smooth out the varying instantaneous HTTP/ TCP reception rates Expected segment fetch time for DASH The spare network bandwidth and network congestion can be identified by comparing ESFT with the measured segment fetch time. ESFT is determined by the following

10 10 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] Fig. 4. An example of ideal parallel segment fetching method to demonstrate the relationship between ISFT and MSD. two factors. One factor is the media segment duration multiplied by the number of parallel HTTP threads which represents the ideal time to fetch each segment ðisftþ. The other factor is the available time to fetch the next segment to maintain the target buffered media time ðtbmtþ. TBMT is used to smooth the short-term transmission variation. The available time to fetch the next segment is denoted as the remaining segment fetch time ðrsftþ. ISFT represents the intrinsic segment fetch time which is determined as the period of time to fetch each segment when the media bitrate of each segment exactly matches the capacity of the network resources. The intrinsic segment fetch time relies on the segment fetching method as ISFT is different with the different n u p. We will give an example to demonstrate how the segment fetch method affects ISFT after presenting the equation to calculate it. While RSFT represents the real-time buffering status of DASH ISFT RSFT and ESFT are computed as follows in the following paragraphs. ISFT is determined as ISFT ¼ MSDn u p ð4þ As discussed in Section 4 the serial segment fetching method can be regarded as a special case of the parallel segment fetching method wherein n u p equals 1. Fig. 4 shows an example of demonstrating a scenario in which the media bitrate exactly matches the network capacity. In Fig. 4 the purple and dark blue bars and lines represent parallel HTTP threads as in Fig. 3. In addition the playback timestamp of the first frame of each segment is labeled as ts i the blue bars and line denote playback duration of each segment and playback timestamp axis which also applies to the following figures depicting the parallel segment fetching method. In the ideal scenario each segment is fetched using exactly ISFT which equals to the media segment duration multiplied by n u p e.g. 2 in this example. It should be noticed that the ideal scenario does not assume an instantaneous reception rate matching the bandwidth but assumes only an average reception rate matching the bandwidth. In practice the above discussed ideal scenario is hard to achieve because of the following reasons. First there is typically a slight mismatch between the optimum media bitrates and the network capacity. Second the network resources and throughput are dynamically changing. To smooth short-term HTTP/TCP throughput variation it is beneficial to maintain a TBMT indicating a desired latency between the time instant to receive the last byte of a segment to the playback of the first media frame of a segment. The RSFT depends on the current playback timestamp ðts 0 Þ at the time instant of requesting the next segment the timestamp of the first frame of the next segment ðts ns Þ and TBMT. Hence RSFT can be calculated as RSFT ¼ ts ns ts 0 TBMT ð5þ Based on Eqs. (4) and (5) finally ESFT is determined as ESFT ¼ minðisftrsftþ ð6þ In this equation ESFT represents the optimum segment fetch time which is used in the rate adaptation metrics. ISFT will be constant with a specific n u p but RSFT will vary dynamically. The reason for using ISFT in the proposed rate adaptation metric is to prevent possible fluctuation in representation level. RSFT is another representation of the buffered media time. Hence if only RSFT were used to determine ESFT then the representation level might fail to converge to an optimum representation level as discussed in Section 2. A large RSFT might be a consequence of a cumulative increase of buffered media time. For example buffered media time may be accumulated due to a step-wise representation-level switch-up strategy. Therefore RSFT might not provide an optimal value for ESFT. In contrast we propose to determine ESFT as the minimum of ISFT and RSFT. The former represents the intrinsic segment fetching time determined by the segment fetching method and the latter represents the real-time buffering status. Hence the proposed method can efficiently facilitate the convergence property but also prevent buffer underflow. Fig. 5 shows an example of determining ESFT for the parallel segment fetching method. In this example it is assumed that n u p is equal to 2 and TBMT is equal to MSD. The last byte of segment #i 2 is assumed to be received

11 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 11 Fig. 5. Example of determining ESFT for the parallel segment fetching method. at the client later than expected as in the ideal scenario depicted in Fig. 4. Consequently the HTTP GET request for segment #i is also sent later than in the ideal scenario. So RSFT is determined using Eq. (5) and is lower than ISFT. According to Eq. (6) ESFT equals to RSFT as shown in Fig. 5 which is used in the rate adaptation metrics and the rate adaptation algorithms as discussed in the following sections SFT rate adaptation metric ESFT represents the optimum duration for fetching a segment. If the measured segment fetch time ðsftþ is equal to ESFT calculated before sending the HTTP GET segment request then each segment is fetched using the optimum segment fetch time and TBMT amount of media data will be maintained at the client buffer. Hence the encoded media bitrate of the current representation matches the network capacity. If SFT is larger than the calculated ESFT then it means that the average TCP throughput for fetching the segment is lower than the bitrate of the current representation level. Otherwise (if SFT is lower than the calculated ESFT) the smoothed TCP throughput is higher than the bitrate of the current representation. The last case can occur in DASH because the TCP sender transmits the available data at the highest possible rate allowed by the TCP congestion control and avoidance algorithm. Hence a ratio of ESFT and the measured SFT is used as the metric denoted as SFTM to detect congestion and to probe the spare network capacity. SFTM ¼ ESFT=SFT where ESFT is calculated as in Eq. (6) each time before sending an HTTP GET request for the next segment and SFT is measured after a segment is fetched. The proposed rate adaptation metric can be used to identify if the media bitrate matches the available network capacity as our metric measures the reception status in a given duration to smooth out short-term variations caused by the TCP congestion control. ð7þ 5.3. PFT rate adaptation metric As PFT measures the time spent to fetch portions of a segment instead of the whole segment PFT should be compared with the expected portions fetch time ðepftþ instead of the ESFT where the EPFT denotes the optimum time to fetch the corresponding portions which have been received by the DASH client. Therefore PFTM is computed as follows: PFTM ¼ðESFTUMPDÞ=ðPFTUMSDÞ ð8þ PFTM rate adaptation is useful in the rate adaptation of the parallel segment fetching method. In the parallel segment fetching method an HTTP GET request can be sent when the previously requested segments have not been completely received at the DASH client. In such a scenario PFT represents the latest reception status and reveals the current network situation. On the other hand the measuring duration should be larger than a certain duration to smooth out short-term TCP throughput variations. Several methods of partitioning a segment into sub-segments dynamically and requesting parallel subsegments instead of requesting parallel segments were discussed in [19]. In this paper we focus on the rate adaptation and PFT used in one of the rate adaptation metrics. One advantage of SFTM and PFTM is that our methods do not require information from the transport layer (TCP layer). Our method uses the reception status at the application layer but does not rely on the TCP throughput calculation models. In order to use the TCP throughput calculation equations the packet loss rates RTT or the current congestion window are required; however such information is not available at the application layer. In addition if the current congestion window is included in the TCP throughput calculation equation then the TCP throughput equation basically represents the instantaneous transmission rate instead of the smoothed TCP throughput. As it is well known the instantaneous TCP throughput is varying and cannot be directly used in the rate adaptation to estimate the end-to-end networks capacity. In contrast our rate adaptation metrics can smooth out the varying TCP throughput rate by measuring a certain duration of the reception status. Therefore

12 12 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] our rate adaptation metrics can be applied for the rate adaptation of DASH which is operated in the application layer. 6. Proposed rate adaptation algorithm for the serial segment fetching method In this section we propose a rate adaptation algorithm for the serial segment fetching method based on SFTM presented in Section 5. In the serial segment fetching method the HTTP GET next segment request is sent after receiving the current segment. The rate adaptation algorithm occurs each time after receiving a media segment and before requesting the next segment to determine the representation level of the next segment. In the serial segment fetching method since ISFT is equal to MSD Eq. (7) can be expressed according to Eq. (6) as SFTM ¼ MINðMSDRSFTÞ=SFT ð9þ In our previous work [10] a simplified rate adaptation metric was presented which was set to the ratio of MSD and SFT. The rate adaptation metric in [10] only considers ISFT of the serial segment fetching method but does not take into account RSFT. In real-time DASH different RSFTs of different DASH clients will cause unfairness especially to those short lived and lately joined clients. In the beginning of DASH if RSFT calculated by Eq. (5) is smaller than the priority RSFT for a serial segment fetching method ðrsft sþ in Eq. (10) then RSFT is replaced by RSFT s. Once RSFT calculated by Eq. (5) is larger than RSFT s RSFT will always be set as Eq. (5). RSFT s ¼ rumsd ð10þ In this paper r is selected as RSFT s is especially important for DASH clients whose initial buffered media time is much lower than the ISFT. The priority RSFT has a larger impact on the parallel segment fetch method compared to the serial segment fetching method since the former occupies the larger ISFT compared to the later. The priority RSFT will be further discussed in Section 7. The rate adaptation deploys a step-wise switch-up and a multi-step based switch-down method to change the representation level. The condition of switch-up/down and the corresponding operations are described as follows: Switch-up: It takes place if the following condition holds: SFTM 41þe u ð11þ In Eq. (11) the left term represents the metric to detect spare network bandwidth and the right term denotes the condition to switch-up to the next higher representation level. e u is determined according to the maximum bitrates jump-up ratio of all representation levels e max u and bitrates jump-up ratio of the current representation level (e c u ). emax u is calculated as e max u ¼ maxfðb ri þ 1 b ri Þ=b ri 8i ¼½01:::N 1Šg ð12þ e max u provides a conservative switch-up factor which improves the convergence property of the proposed rate adaptation algorithm. And e c u is calculated as e c u ¼ðb b Þ=b rc þ 1 rc rc e c u ð13þ provides a more aggressive switch-up factor compared to e max u. Since e c u is only determined by the bitrate jump ratio of the current representation level and e max u is decided by the maximum of bitrate jump ratios of all representation levels. To reach a compromise between the convergence of the rate adaptation algorithm and the achievable media bitrates e u is calculated as e u ¼ minfe max u 2e c u g ð14þ Eq. (14) not only improves the convergence and achievable media bitrate but also makes the rate adaptation algorithm performance independent of the bitrate distribution of the representation levels. For example if e max u is much larger than e c u then e u can be determined as 2e c u which enables the rate adaptation algorithm to switch-up the representation level to the optimum level. Otherwise (e max u is smaller than 2e c u ) e u will be decided by e max u for improving the convergence. In the case of a decision to switch-up the rate adaptation algorithm selects the next higher representation level. The reason for using a conservative step-wise switch-up strategy is to prevent playback interruption which might occur in case of aggressive switch-up operations. During a step-wise switch-up moreover buffered media time can rise to a more safety level to prevent buffer underflow due to a sudden bandwidth decrease. So the initial buffering time spent to buffer the media time to reach its minimum can be set as a relatively small value to reduce the delay before starting playback. Switch-down: It will be performed if the following inequality holds SFTM o1 e d ð15þ The left term of Eq. (15) compares the selected media bitrate to the current bandwidth in terms of ratio of the optimum segment fetch time to the actual segment fetch time while the right term represents the condition to switch-down to a lower representation level. To identify e d the minimum bitrate switch-down ratio of all representation levels ðe min d Þ and bitrates switch-down ratio of the current representation level ðe c dþ are used. emin d is calculated as e min d ¼ minfðb ri b ri 1 Þ=b ri 8i ¼½01:::N 1Šg ð16þ e c d is calculated as e c d ¼ðb b rc r c 1 Þ=b rc ð17þ Then e d is calculated as e d ¼ maxf2e min d e c d g ð18þ In case of congestion SFT will be much larger than ESFT i.e. RAM is lower than 1. Hence Eq. (18) selects a switch down factor at least larger than a specific value such as 2e min d which is used in this paper. When e c d is larger than 2emin d wesete d as e c d. Eq. (18) enables the proposed rate adaptation algorithm to provide an advanced convergence property. For a rate adaptation algorithm of DASH providing an advanced convergence property is one of most challenging task since the switch-up and switch-down behaviors from different DASH

13 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 13 clients will affect other clients. In this paper we propose Eqs. (14) and (18) to provide a conservative switch-up and switch-down for improving the convergence property. The proposed rate adaptation algorithm of the serial segment fetching method is described as the following pseudocode. Proposed Algorithm 1. Rate adaptation of the serial segment fetching method operated at the client. Initialize the current representation level i as 0 Send HTTP GET the first segment request immediately While User has not stopped the current DASH and more segments are available for fetching Wait for an event If (Received a segment and t id r0) or (timeout to request the next segment) If switch up condition meets Set representation level to the next higher level Else if switch down condition meets Representation level i is determined as the first representation r i to meet the inequality (19) Else Representation level remains unchanged Send HTTP GET the next segment request immediately If Received a segment and t id 40 Idle t id period before sending HTTP GET the next segment request End while To support the conservative switch-down a client buffers enough amount of media time to recover from a severe bandwidth decrease. To accumulate media time this paper deploys an additive increase and multiplicative decrease strategy to adapt the representation level. In the switch-down an aggressive switch-down will be performed. The selected representation level is determined to be the first representation (in descending order) with level r i to meet b ri osftmub rc ð19þ The idle time calculation algorithm is deployed before sending the next GET request in order to prevent client buffer overflow in a similar manner as Eq. (3) presented in the parallel segment fetching method. 7. Proposed rate adaptation algorithm for parallel segment fetching method In this section we present a novel rate adaptation algorithm for the parallel segment fetching method based on the proposed rate adaptation metrics i.e. SFTM and PFTM. In this section SFTM and PFTM are uniformly denoted as rate adaptation metric RAM. The rate adaptation algorithm also deploys a step-wise switch-up and an aggressive switch-down method which are determined according to the following two cases. Switch-up: It takes place if all RAM t of the parallel received segments meet the following condition RAM t 41þe u 8t ¼½t 0 t 0 þ1:::t 0 þw 1Š ð20þ Fig. 6. An example of demonstrating unfairness between different clients due to the different RSFTs: (a) first client playback timestamp of each segment and the ESFT to fetch the segment #i and (b) second client playback timestamp of each segment and the ESFT to fetch the segment #j.

14 14 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] where the left term represents themetrictodetectconges- tion and the right term denotes the condition to switch up to the next higher representation level. In the above switch-up condition a sliding window is used to compare the latest w number of RAM t with the switch-up threshold to decide whether or not to perform a switch-up. And the rate adaptation index (t) startsfromt 0 instead of 0 to set a minimum received potion media duration for each RAM t. RAM t can be used as the rate adaptation metric if the ratio of the received portion duration and the segment duration is larger than a predefined threshold (d) orpft is larger than ISFT as MPD t 0 ¼ arg k minðkþ MSD Zd or MPD k 4ISFT 8k ¼½01:::n p 1Š ð21þ The reason of setting t 0 is to smooth the short-term TCP transmission variation. d is set as 0.7 in this paper. e u is calculated as in Eq. (14). Switch-down: It will be performed if the RAM t0 meets the following condition: RAM t0 o1 e d ð22þ t 0 is the starting rate adaptation index determined as Eq. (21). e d is calculated as in Eq. (18) which provides a conservative switch-down performance since e d is set as at least larger than 2e min d. In the switch-down an aggressive switch-down will be performed. The selected representation level is determined to be the first representation (in descending order) with level r i to meet b ri oram t0 b rc ð23þ The fairness between different clients is one of the important factors to evaluate the efficiency of a rate adaptation algorithm. Fig. 6. shows an example demonstrating the occurrence of an unfairness between the two clients because of the sustained different ESFTs. Fig. 6(a) and (b) depict the ESFTs of the first and the second clients respectively. In Fig. 6 the same term and definition of the symbols are used as in Fig. 4 to depict the playback timestamps of the first frame of each segment the playback durations of each segment and the playback timestamp axes. It is assumed that the first client starts DASH earlier than the second client. In this example we assume that the spare network bandwidth is available when the first client starts DASH and the network bandwidth is scarce when the second client starts DASH. In Fig. 6(a) the first client reaches a relatively higher representation level 5 and buffers a large amount of media time at time instant t_req i using the spare networks. The second client in Fig. 6 (b) however stays at a lower representation level 1 and only buffers a small amount of media time at time instant t_req j as the networks enters a saturated status when the second client starts DASH. The different buffered media time results in different RSFTs which further causes different ESFTs as demonstrated in Fig. 6. To the first client shown in Fig. 6(a) the current playing segment is segment i 5 when the first client requests segment i at time instant t_req i.inthisexampleisft equals to two times MSD as n u p is set as 2 and TBMT equals MSD. Since RSFT covers more than two segment durations it is larger than ISFT. SoESFT needed to fetch segment i is equal to ISFT according to Eq. (6). For the second client shown in Fig. 6(b) the current playing segment is j 3 when the second client requests segment j at time instant t_req j. Under the same TBMT with the first client RSFT of the second client is less than ISFT. Thus ESFT required to fetch segment j is equal to RSFT according to Eq. (6) for the second client. As networks undergo a saturated status it might be difficult for the second client to accumulate buffered media time and RSFT may be sustainably lower than ISFT. In real-time DASH different ESFTs of different clients will cause different achievable media bitrates since ESFT determines RAM SFTM and PFTM according to Eqs. (7) and (8). To solve the problem of unfairness caused by different RSFTs of different clients this paper presents a method of deploying the priority RSFT for parallel segment fetching method ( RSFT p) in the beginning of DASH. If RSFT calculated by Eq. (5) is smaller than RSFT p calculated by Eq. (24) then RSFT is replaced by RSFT p. Once RSFT calculated by Eq. (5) is larger than RSFT p RSFT will be calculated as in Eq. (5). RSFT p ¼ð1þðn u p 1Þ=2ÞMSD ð24þ RSFT p allows newly started DASH clients to successfully switch-up to the appropriate medium representation level. The rate adaptation algorithm of the parallel segment fetching method is shown in the following pseudocode which operates as the rate adaptation module in Algorithm 1. Proposed Algorithm 2. Rate adaptation of the parallel segment fetching method operated at DASH client. If switch-up condition is met Set representation level to the next higher level Else if switch-down condition is met Representation level i is determined with the first representation r i to meet Eq. (23) Else Representation level remains unchanged End 8. Simulation results The proposed rate adaptation algorithms for the serial and the parallel segment fetching methods in CDN are implemented in ns2 [25]. In ns2 the proxy cache module consists of the standard proxy cache the client and the server which operate on HTTP/TCP protocols. Recently Müller and Timmerer [26] presented a reference software for DASH so that rate adaptation algorithms can be implemented in the reference software for evaluation. The bandwidth dynamic and packets loss are emulated using the traffic shaping tools operated on the rooter or at the end user device. However emulating the distributed networks such as CDN is not an easy task. For emulating CDN first a distributed network topology should be created. For evaluating the dynamics of the distributed

15 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 15 network resources each edge server should be equipped with the traffic shape tool. Second it is complex to specify the transporting pass for traffics in the distributed networks e.g. determining the route between the DASH server and the DASH client for each segment. In contrast all of the above issues are easy in ns2. Therefore ns2 is a good choice for evaluating the rate adaptation method for DASH in CDN. In this paper we evaluate the rate adaptation algorithm in the aspects of (a) convergence in the requesting representation level indicating the media bitrates (b) fairness between different clients which compete for the edge server bandwidth (c) achievable media bitrate (d) rate adaptation speed and (e) buffer underflow frequency. To our knowledge few rate adaptation algorithms are published in the literature. Several HTTP streaming players exist such as Microsoft smooth streaming [5] and Adobe OSMF player [6] have their own internal rate adaptation techniques; however they are not public. Hence it is not feasible to compare the proposed rate adaptation algorithm with those players. Although we cannot implement the specific rate adaptation algorithm it is possible to observe the behavior of the commercial players. As mentioned in the related work Section of this paper Akhshabi et al. [15] reported the experimental results and the evaluation focusing on the evolution of requesting media bitrates over time using commercial players. In Section 8.5 of this paper samples of the representation level evolution are reported for 10 different clients that compete for the bandwidth of the edge servers. The proposed rate adaptation algorithms are compared with the Smooth streaming reported in [15]. Reporting samples of evolution of the representation levels has limitation in the coverage of reporting results considering the page limitation. For example when the DASH client group size was increased to and 30 or even more it is not feasible to observe each client evolution process. To conduct and report extensive simulation results the distribution of the representation levels of DASH clients are reported. The histogram of the representations level which indicates the occurrence frequency of each representation level is an efficient tool to evaluate the distribution of the representation levels. Reporting each client histogram will also take many pages with the large number of simulated clients. Hence each histogram is represented as the combination of two values which capture the features of the histogram and the comprehensive results and details are reported in Section 8.2. The mean of the average media bitrates is reported to evaluate the achievable media bitrates of the overall clients and the buffer underflow frequency is reported in Sections 8.3 and 8.4 respectively. Finally samples of the representation level evolutions are reported in Section 8.5. Table 2 Two sets of bandwidth settings. Bandwidth settings Links #1 Links #2 Links #3 Bandwidth setting #1 (Mbits/s) Bandwidth setting #2 (Mbits/s) Table 3 Group size with the bandwidth setting #1. Bandwidth settings Minimum Maximum Step Bandwidth setting # ExpR-0 ExpS-8 ES-0 ExpS-0 ExpS-1 Link #5 CDNS-0 ExpR-1 ExpS-9 ExpR-8... ExpR- 11 DASH client-0 ExpR-0 ExpS-2 Link #1 Link #2 ES-1 Link #4 DASH server... CDNS-4 ExpR-6... ES-6 ExpS- 14 Link #3 ExpR- 12 Internet connector... ExpR- 15 DASH client-1... DASH client-n DASH server Link #5 Link #1 ExpS-0 CDNS- 0 ES-0 Link #2 Link #3 ExpR-2 Internet connector Link #4 DASH client-0 DASH client-1 ExpS-6 ExpS-7 ES-7 ExpS-1 ES-1 ExpR-3... DASH client-8 ExpR-7 ExpS-15 ExpR-1 ExpS-3 Fig. 7. Network topology of the simulation. (a) Topology #1 (b) Topology #2.

16 16 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 8.1. Simulation topology and setting Fig. 7 shows the network topologies used in the simulation. In the first topology Fig. 7(a) two layers of media distribution networks are deployed including four CDN servers (CDNS-0 CDNS-1 CDNS-2 and CDNS-3) and eight edge servers (ES-0 ES-1 y ES-7). In topology #2 two layers of media distribution networks are deployed including one CDN server (CDNS-0) and two edge servers(es-0 and ES-1). As shown in Fig. 7 the media segments are delivered from HTTP streaming server to CDN servers edge severs the Internet connector and finally to DASH clients. The edge server fetches the segment from the connected upper layer CDN server which then fetches the segment from the DASH server. The DASH server contains all representations of the media clip. In this paper all segments are fetched from the DASH server. Caching scenario is not considered in this paper. To simulate the above scenario different clients request the different media clips hence all segments have not been cached by the edge servers when they receive a HTTP GET segment request. Table 2 shows the bandwidth settings of links #1 links #2 and links #3 in Fig. 7. The bandwidths in links #4 and links #5 are set as 2.0 Mbits/s. All link delays are set to 2 ms. Topology #1 in Fig. 7(a) and the bandwidth setting #1 are used to report the simulation results of Sections and 8.4. Topology #2 in Fig. 7(b) and bandwidth setting #2 are used to report the representation level evolution in Section 8.5. To simulate practical congestion in real networks and evaluate the rate adaptation algorithm multiple DASH clients are simulated which compete for the bandwidths of the distributed edge servers in CDN. The clients are classified into two groups. For the first group clients start the DASH randomly in the time interval between 0 s and 10 s and end at 1200 s. For the second group clients start the DASH randomly in the time interval s and stop at 800 s. By increasing or reducing the number of clients denoted as group size the rate adaptation efficiency can be evaluated for different levels of congestion. Here the group size is the sum of the client s number in the two groups. Table 3 shows the minimum group size maximum group size and group size jump step for the bandwidth setting #1. As shown in Table 3 the group size is much larger than the number of edge servers. Hence multiple clients will compete for the bandwidth of the links between the CDN server and the edge server and the links between the edge server and the Internet connector in Fig. 7. Forthe bandwidth setting #1 the simulation runs multiple times with different group sizes. The numbers of clients in the first and the second group are equal. For the first DASH client group the simulation time is divided into three periods to report simulation results. The three periods include s i.e. before starting the second client group s i.e. when the second client group compete for the bandwidth of the distributed networks and s i.e. after ending the second client group. For the second client group the simulation results are reported in the period starting from 400 s and ending at 800 s. To simulate the dynamics of the Internet background traffic is added to each link between the CDN server and the edge server and between the edge server and the Internet connector by attaching the exponential traffic senders (ExpS-id) and the corresponding receivers (ExpRid). Specifically the exponential traffic senders are attached to the CDN servers (CDNS-id) and edge server (ES-id). The corresponding exponential receivers are attached to the edge servers (ES-id) and the Internet connector. Both operating and idling periods of the exponential traffic are set to 0.5 and the average bitrate of exponential traffics during the operating period is set to 0.2 Mbits/s. The exponential traffics start from 0 s and end at 1200 s. The edge server selection method and the frequency of switching the edge server affect the achievable media bitrates of the serial and the parallel segment fetching methods. When a simple edge server selection method such as round-robin strategy is used the frequent switching of the edge server balances the edge server loads better than the infrequent switching and will more efficiently utilizes the bandwidths of distributed edge servers. Hence the achievable media bitrate of serial and parallel segment fetching methods can be improved. As discussed in Section 3 the proximity based edge server selection technique causes infrequent request redirection and the short TTL based load balancing algorithm results in frequent edge server switching. Recall that the edge server selection method and determination of the frequencies of switching edge server are out of the scope of this paper. Therefore the different frequencies of the switching edge server are simulated in order to evaluate the proposed rate adaptation algorithms for the serial and the parallel segment fetching methods with respect to the different switching frequencies. In the simulation the frequencies of 1 and 0.25 are used for switching edge servers indicating switching edge servers every one segments and four segments respectively. In addition clients randomly select the initial edge server from the set of edge servers to fetch the first segment. Under a specific switching frequency round-robin strategy is used to switch the edge server. In our simulations TBMT and the initial buffered media time are set to 20 s. For achieving adaptive HTTP streaming the server provides ten sets of representations for clients to adapt the media bitrates wherein the media bitrates include and 1408 Kbits/s and the corresponding representation level starts from 0 and ends at 9 respectively. Bitrate changing step was set to increase as representation level reaches to a higher level in order to achieve an equivalent quality improvement each time bitrate switches-up Distribution of representation level The histogram of the representation level r i is an efficient metric for evaluating the rate adaptation algorithm since the histogram indicates the frequency distribution of the inputs. In the histogram calculation the input is each segment representation level x ¼fx 0 x 1...x H 1 g and the output includes a vector of the occurrence frequency of all possible representation levels

17 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 17 Fig. 8. An example of representation level histograms of two clients and corresponding (s 2 f s2 l ): (a) first client histogram ( ) and (b) second client histogram ( ). Fig. 9. Representation level distribution of each client using the proposed rate adaptation algorithm for the serial segment fetching method under the edge server switching frequency 1: (a) s (b) s and (c) s. Fig. 10. Representation level distribution of each client using the proposed rate adaptation algorithm for the parallel segment fetching method under the edge server switching frequency 1: (a) s (b) s and (c) s. c r ¼fc r0 c r1...c rm 1 g and a vector of the distinct representation level location br ¼fbr 0 br 1...br N 1 g. In the former vector the vector size of occurrence frequency M is equal to the number of all possible representation levels provided by the content provider and the occurrence count is set as zero for a representation level that has not occurred. In the later vector if a representation occurrence count is larger than or equal to a constant threshold y then the corresponding representation level is regarded as a distinct representation level and the corresponding representation level value denotes the distinct representation level location. In this paper y is set as 5. Such a distinct representation level locations ðbr k Þ forms a vector of distinct representation level locations denoted as br. br ¼fbr 0 br 1... d r N 1 g¼arg rk ðc rk Zy k ¼ 01...M 1Þ ð25þ where N denotes the size of distinct representation level location which is smaller than or equal to M. The variance is calculated over c r and br to obtain the variance of the occurrence frequency of all possible representation levels (s 2 f ) and the variance of distinct

18 18 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] Fig. 11. Representation level distribution of each client using the proposed rate adaptation algorithm for the serial segment fetching method under the edge server switching frequency 0.25: (a) s (b) s and (c) s. Fig. 12. Representation level distribution of each client based on the proposed rate adaptation algorithm for the parallel segment fetching method under the edge server switching frequency 0.25: (a) s (b) s and (c) s. representation level location (s 2 l ) as follows: P M 1 s 2 i ¼ 0 f ¼ ðc r i c r Þ 2 M ð26þ P N 1 s 2 i ¼ 0 l ¼ ð br ı br Þ 2 ð27þ N where c r denotes the mean of the occurrence frequency of all possible representation levels and br denotes the mean of distinct representation level locations respectively. A large s 2 f and a smaller s 2 l indicate a superior convergence property and vice versa. Fig. 8 shows an example of calculating s 2 f and s 2 l for two given histograms of representation levels and comparing the convergence property of the two histograms by comparing the combination of s 2 f and s 2 l denoted as (s2 f s2 l ) of the two histograms. As shown in Fig. 8 c r and br of the first and the second histogram are f g f235g and f g f45g respectively. As can be observed from Fig. 8 the second histogram represents a superior convergence compared to the first histogram because of the following two reasons. First the shape of the second histogram is sharper than that of the first histogram. Second the distinct representation levels are more focused in the second histogram compared to the first histogram. According to Eqs. (26) and (27) (s 2 f s2 l ) of the first and the second histogram are ( ) and ( ) respectively. The second histogram gives a larger s 2 f and a smaller s 2 l compared to the first histogram. As mentioned before a larger s 2 f and a smaller s 2 l indicate a superior convergence hence the second histogram represents a more advanced convergence which is consistent with our observation. Figs. 9 and 10 plot the (s 2 f s2 l ) of each client for the rate adaptation algorithms of the serial and the parallel segment fetching methods respectively in the interval of s s s with the edge server switching frequency 1. Figs. 11 and 12 show the representation level distributions similar to Figs. 9 and 10 but with the edge server switching frequency The horizontal and vertical axes represent s 2 l and s 2 f respectively. The dots distributed in the left-up corner of the figure indicate better convergence compared to the dots distributed in the right-bottom corner of the figure according to the above mentioned rule that a larger s 2 f and a smaller s 2 l indicate better convergence performance. Four sets of group sizes from group size 8 to 32 are reported. Comparing the convergence properties of the three periods in Figs. 9 and 10 the best results are given in the period of s medium results are in period s and the worst results are shown in period s for the proposed rate adaptation algorithms. The reasons are as following. In the period of s less clients compete for the bandwidth compared to the period of s and the starting representation level is larger than 0. In the period of s the representation level starts from 0. In the period of s more clients compete for the bandwidth compared to the other periods. In addition the convergence property of the rate

19 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] 19 adaptation algorithms depends on the group size according to the simulation results in Figs. 9 and 10. When the group size is 8 or 16 the rate adaptation algorithms provide an improved convergence. The convergence level is still acceptable for the group size 24. But it becomes worse when the group size was increased to 32 which needs to be improved further. However the above results already show that the proposed rate adaptation algorithm is superior to the behavior of the Smooth streaming player wherein the two players compete for the bandwidth of a common link which was reported in section 6 of [15]. According to the evaluation reported in [15] the second player which started later than the first player failed to switch-up to the appropriate level and kept oscillating between some lower representation levels due to the competition from only two players. Improved convergence is reported in Figs. 9 and 10 with the group size from 8 to 24 compared to the results and evaluation in [15] wherein the group size was 2. Providing improved convergence for a large group of DASH clients in CDN is still an open and challenging task because of the following reasons. In the scenario of multiple clients competing for the bandwidth in the distributed edge servers the number of clients may be dynamically changing which causes variation in the sharable bandwidth. Furthermore the rate adaptation of one client might cause a change in the rate adaptation of other clients because of the varying sharable bandwidth. For example when several clients compete for the bandwidth in a common link a switch-up behavior of a client may cause the reduced sharable bandwidth for other clients while a switch-down behavior of a client may result in the increased sharable bandwidth in the other clients. Such a relation between different clients is called inter-client effect in this paper. The inter-clients affection will increase along with an increase in the group size. Moreover the edge server selection and load balancing algorithm affect the variation in the sharable bandwidth of each client and hence affect the convergence property of the rate adaptation algorithm of DASH. Comparing the rate adaptation algorithms for the serial and parallel segment fetching methods shown in Figs Figs respectively the simulation results show that the rate adaptation algorithm for the serial segment fetching method slightly outperforms the rate adaptation algorithm for the parallel segment fetching Fig. 13. Mean of average bitrates of the first group with edge server switching frequency 1: (a) the first group in s and (b) the first group in s. Fig. 14. Mean of average bitrates of two groups with edge server switching frequency 1: (a) the first group in s and (b) the second group in s.

20 20 C. Liu et al. / Signal Processing: Image Communication ] (]]]]) ]]] ]]] Fig. 15. Mean of average bitrates of the first group with edge server switching frequency 0.25: (a) the first group in s and (b) the first group in s. Fig. 16. Mean of average bitrates of two groups with edge server switching frequency 0.25: (a) the first group in s and (b) the second group in s. method in the terms of convergence in the representation level. Comparing the rate adaptation performance with respect to the different edge server switching frequency no significant difference is observed Average bitrates with varying group size Table 4 Buffer underflow frequency. Group size Buffer underflow frequency Edge server switching frequency 1 Edge server switching frequency 0.25 The mean of the average media bitrates for the different group sizes are reported with the bandwidth setting #1 for analyzing the rate adaptation algorithms with the respect to the achievable media bitrates under the conditions of the different crowded networks. In Figs the y and x axes denote the mean of the average media bitrates of each client group and the group size respectively. And the Serial-DASH-G1/2 and the Parallel- DASH-G1/2 denote the rate adaptation results for the serial and parallel segment fetching methods in the first/second group receptively. Fig. 13(a) (b) shows the mean of the media bitrates of the first group in the period of s and s respectively. In both periods the distributed networks bandwidths are only used by the clients of the first group without competition from the second group. So the number of active clients in the period of s and RA for the serial RA for the parallel RA for the serial RA for the parallel s is half of the group size. Fig. 14(a) (b) shows the mean of the media bitrates of the first and the second group in the period of s. In this period the distributed bandwidths are shared by the clients of the two groups so the number of active clients is equal to the group size. Comparing the rate adaptation algorithms for the serial and the parallel segment fetching methods the mean of the media bitrates of the rate adaptation algorithm for the parallel segment fetching method is higher than that of the